Predicting Crime without the Pre-Cogs?
Crime in Chicago is difficult to predict. Hyde Park is remarkably safer than when I was a student here a decade ago (and in fact is one of the city's safest neighborhoods), but this past week saw a troubling and sad increase in violence in other parts of the city. The Mayor and police attributed some of the increase to the warm weather, but policy makers and citizens seemed surprised at the spike, which goes against national and regional trends. Predicting when crime will occur and the impact that different policing and punishment policies will have on crime is a tricky business. So how are policy makers supposed to decide how many resources to deploy, where to deploy them, and what the impact of, say, a moratorium on the death penalty (as we had here in Illinois recently) will have on crime?
Fiction offers us one possibility. In the movie version of Philip Dick's story "Minority Report," a police officer played by Tom Cruise leads a Division of Pre-Crime that with the use of three semi-conscious "pre-cogs" can predict certain crimes before they happen. When the pre-cogs "see" a murder a few minutes in the future, the police rush to arrest the perpetrator before the act is committed. The movie, which is fascinating and brilliantly done, raises a host of legal issues, not the least of which is whether punishment of mens rea without actus reus is sensible public policy. But we don't have pre-cogs. (At least not that I know of.)
In the absence of pre-cognitive superbeings and Tom Cruise, what are the alternatives? There is a vibrant academic literature on predicting crime, with models of various types offered as the best way of estimating future crime rates. Many of these involve mapping software, which plots the past in the hopes of extrapolating to the future. Police use some of these techniques, but most are very crude, using things like weather or the location of liquor stores as "hot spots" to estimate crime rates. Police, we are told by them, also use experience and gut instinct. All of the various methods, whether formal models or inside the head of the commissioner of police, are deployed in haphazard and isolated ways.
In a paper recently posted to SSRN, co-authors Justin Wolfers (Wharton) and Eric Zitzewitz (Dartmouth) and I offer an alternative that is designed to extract from the "market" estimates that aggregate and incorporate all readily available information about future crime rates under various policy choices. Here is the abstract:
Prediction markets have been proposed
for a variety of public policy purposes, but no one has considered
their application in perhaps the most obvious policy area: crime. This
paper proposes and examines the use of prediction markets to forecast
crime rates and the impact on crime from changes to crime policy, such
as resource allocation, policing strategies, sentencing,
post-conviction treatment, and so on. We make several contributions to
the prediction markets and crime forecasting literature.
First, we argue that prediction markets are especially useful in crime rate forecasting and criminal policy analysis, because information relevant to decisionmakers is voluminous, dispersed, and difficult to process efficiently. After surveying the current forecasting practices and techniques, we examine the use of standard prediction markets - such as those being used to predict everything from the weather to political elections to flu outbreaks - as a method of forecasting crime rates of various kinds.
Second, we introduce some theoretical improvements to existing prediction markets that are designed to address specific issues that arise in policy-making applications, such as crime rate forecasting. Specifically, we develop the idea of prediction market event studies that can be used to test the influence of policy changes, both real and hypothetical, on crime rates. Given the high costs of changing policies, say issuing a moratorium on the death penalty or lowering mandatory minimum sentences for certain crimes, these markets provide a useful tool for policy makers operating under uncertainty.
These event studies and the other policy markets we propose face a big hurdle, however, because predictions about the future embed assumptions about the very policy choices they are designed to measure. We offer a method by which policy makers can interpret market forecasts in a way that isolates or unpacks underlying crime factors from expected policy responses, even when the responses are dependent on the crime factors.
Finally, we discuss some practical issues about designing these markets, such as how to ensure liquidity, how to structure contracts, and the optimal market scope. We conclude with a modest proposal for experimenting with markets in this policy area.